| name | oat-wave-execute |
| version | 1.7.1 |
| description | Use when executing a wave of external implementation plans as a wrapper OAT project — scaffolding, drift refresh, parallel worktree groups, briefs, gates, merge choreography, and closeout. |
| argument-hint | <wave-id> [plan-names...] (e.g. wave-2 http-listener-before-indexing ...) |
| disable-model-invocation | false |
| user-invocable | true |
| allowed-tools | Read, Write, Edit, Bash, Grep, Glob, Task |
| metadata | {"internal":true} |
Execute a Wave of External Plans
Run one wave of the external-plan program as a wrapper OAT project. This skill
owns the mechanical layer — everything that waves 0–1 proved stable and that
hand-re-derivation repeatedly broke. It deliberately does NOT own judgment.
Provenance: extracted per DR-260713-extract-oat-wave-execute after the source
program's waves 0–1 and hardened across its waves 2–5.
Log every friction/deviation to the project's orchestration-log.md so later
waves inherit evidence, not anecdotes.
Historical DR-* and BL-* slugs in this skill (including the provenance
record above) are evidence citations in the source program's repository; they
are not required artifacts in the consuming repo.
Progress Indicators (User-Facing)
Print this banner once when execution starts:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
OAT ▸ WAVE EXECUTE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
- Before each major boundary, report the current step: preflight, drift refresh,
wrapper scaffold, group dispatch, integration gate, or closeout.
- For parallel groups, report accepted lane IDs once, then summarize outcomes
after all lanes return; do not stream every worker action.
- Surface STOP conditions immediately with the blocked lane, durable evidence,
and exact resume action.
Ownership Boundary
This skill owns (mechanical): integration/phase branch naming, worktree
bootstrap, wave-boundary drift refresh, wrapper-project scaffold (from the bundled
templates), brief templates, gate-prompt template + scaling, merge choreography,
bookkeeping cadence, closeout sequence.
The orchestrator owns (judgment — never delegate to this skill or to workers):
parallel-group composition from live recon evidence, review-finding dispositions,
verification of load-bearing worker claims, merge-order decisions under live drift,
cross-lane synthesis, the end-of-run synthesis, and all user checkpoints.
Standing Rules (violations caused every wave-1 incident)
- Branch naming: integration branch
wave-N-execution; phase branches
wave-N/pNN. Phase branches MUST NOT nest under the integration branch name
(git refs cannot be both leaf and directory).
- Merge commits use a conventional type:
chore(pNN): merge wave-N/pNN — <lane> (review passed).
Bare merge(...) fails commitlint. Keep headers ≤ 100 chars.
- Clean orchestrator tree before group merges — a dirty unrelated file leaves
git merge --no-ff uncommitted and drags the full hook chain into recovery.
When a gate reviewer runs in the primary checkout, it MUST COMMIT its own
artifact.
From a linked worktree, .git points to metadata outside the review sandbox and
that commit can fail silently; the ORCHESTRATOR commits the gate artifact on
the reviewer's behalf. Keep the tree clean around either path (both consumers:
stoa waves 2–3; Orc W1–W4).
- Pre-declare CUMULATIVE churn in every brief: declarations cover everything
landed since each source plan's AUTHORED COMMIT (drift checks compare against
that commit, not the group base), naming files + rough regions. Zero false
drift-STOPs across 32 consecutive briefed lanes through wave 4
(DR-260715-cumulative-churn-manifests).
- Resolve full SHAs with
git rev-parse — never hand-expand short SHAs.
- Gate scope must fit the timeout: bound wrapper-plan gate reviewers to the
wrapper artifacts + the drift-refresh record (not all N source plans), or raise
the gate timeout. On a gate timeout, check the project
reviews/ directory for
a completed artifact carrying the gate runId BEFORE re-running. For larger
waves (~6+ lanes), prefer per-phase gates over one monolithic final gate; if a
final gate must cover the whole wave, scope its prompt to the integration
diff plus the review-chain artifacts, not a re-review of every lane.
Gate dispatch posture (standing rule alongside rules 6 and 8): dispatch
gates in the BACKGROUND by default with a completion watcher because
orchestrator-host foreground ceilings (for example, 600 seconds) are shorter than
legitimate wave-scoped reviews. Use foreground only for demonstrably short scopes.
Rule 8 remains the recovery path; it worked on the one timeout in Orc W2 and was
not needed again in W2–W4.
- Guard every formatter-ignored-file × staged-glob interaction: a
single-glob staged-file task can fail when every matched file is ignored by
the repo's formatter. Audit every glob task and pair canonical-file
exclusions with the staged-task's matching ignore filter (source-program
example: oxfmt exited non-zero for ignored
*.md/*.json, requiring both
.oxfmtrc.json ignorePatterns and .lintstagedrc.mjs filtering).
- Gate timeout diagnostics (the upstream stdin-hang fix landed in oat
0.1.65 — gate children now get
stdin: 'ignore'; the historical
< /dev/null workaround is retired and harmless if still present):
timeout + ZERO output bytes = launch defect (do not spend remediation
attempts re-running unchanged); timeout + a complete artifact = late
completion (recover the artifact by runId per rule 6). Verify the CLI is
≥0.1.65 (oat --version) before dropping the workaround on a new host.
- Verified bookkeeping edits (
DR-260713-bookkeeping-table-mutations):
never mutate plan/implementation status tables with exact-string replacement —
oxfmt re-padding makes it silently no-op. Use anchored regex + a substitution
count assertion + a post-state grep, every time. This discipline caught its
own subsequent no-ops twice in wave 2; treat an assert failure as normal
operation, not an incident. Orc reproduced the failure at wave-close: regex
replacement missed oxfmt padding, and an asserted line-based transform was
required.
- Integration gates after every fan-in: they are the only detector for
cumulative-timing defect classes. Never skip them because every lane passed
independently; the wave-5 embed-teardown defect was caught only after fan-in.
- Fix rounds are APPEND-ONLY: never amend a reviewed SHA; amendment
invalidates stored review verdicts that cite that commit. Every fix-round brief
MUST state that fixes land in an append-only commit. A worker that refuses an
instruction to amend a reviewed SHA is honoring its role contract (Orc p10
precedent).
- Piped DoD/gate verification must preserve the raw failure: run every piped
verification chain under
set -o pipefail, or capture the raw command's exit
code before filtering. In Orc W4, pnpm test | grep returned the filter's zero
while the test run contained one failure.
- Review artifacts are single-writer until committed: an uncommitted review
artifact is exclusively owned by whichever agent is live on it. Orchestrator
dispositions land as immediate commits or wait until every agent touching the
artifact terminates. Do not use lock or timestamp-suffix conventions; they
fragment the review chain that final gates audit (Orc W4 final-gate evidence).
Plus inherited invariants: commit-verification via git log before retrying after
any ambiguous hook outcome; every agent runs the repo's formatter on markdown it
writes; follow the repo's runtime/environment setup before package commands;
repair native-dependency ABI mismatches using the repo's documented rebuild
procedure; rebuild edited workspace-package output before running consumer tests
(stale output mimics real failures). Source-program examples were pnpm format:fix, nvm use, and pnpm rebuild -r better-sqlite3 on
NODE_MODULE_VERSION errors.
Inputs
<wave-id> — e.g. wave-2. Derives project wave-N-execution, branches per
rule 1.
- Lane list: resolve from the live execution-program artifact
(
.oat/repo/reference/external-plans/*-execution-program.md, owned by
oat-wave-program); fall back to the plan indexes' wave hints only when no
program artifact exists. Verify every named plan file exists before
scaffolding; a missing file is a STOP (report, don't guess).
- Concurrency ceiling: default 3 worktrees (operator-set; revisit per wave).
Process
Step 1: Preflight
- From the repo root on up-to-date
main: git checkout -b wave-N-execution.
- Follow the repo's runtime/environment setup, install dependencies, and run
its build/type baseline gates (source-program example:
nvm use && pnpm install --frozen-lockfile, rebuild better-sqlite3 if Node changed, then
pnpm build && pnpm type-check).
- Record
BASE_SHA=$(git rev-parse HEAD).
Step 2: Wave-boundary drift refresh (recon dispatch)
Dispatch ONE economical recon subagent (read-only) over all wave plans:
- Re-run each plan's
## Drift check against BASE_SHA; classify
PASS / MINOR-DRIFT (describe) / STOP (quote the plan's own condition + evidence).
- AUDIT each plan's drift-check FILE COVERAGE against its stated scope: a plan
whose implementation surface includes files its drift-check command omits gets
a wrapper rule-1 addendum (wave-2 p07 precedent — the gap became a mandatory
in-worktree extension and the anticipated conflict never materialized).
- Extract each plan's COMPLETE write-surface list and mechanically intersect
every pair — flag every file written by 2+ plans, however minor (this feeds
group composition; include generated files like
apps/documentation/index.md
when plans run docs gates, and CLI/help/parity files that many lanes touch).
- Return a compact per-plan table + shared-surfaces section. Conclusions only.
A tripped STOP parks that lane at plan time (recorded in the wrapper plan; lane
excluded from groups) — never mid-run. Reconciliation contract: when recon
reveals a stale plan premise, the reconciliation must be NON-NARROWING (WHERE
the work happens may change; WHAT must be true may not), recorded exactly once
in the Drift Refresh Record with pointer-only references elsewhere — a
reconciliation that waives a source-plan requirement is a plan-gate Important
(wave-4 evidence).
Step 3: Scaffold the wrapper project
- Probe the installed scaffold interface first with
oat project new --help.
If --no-commit is present, run
oat project new wave-N-execution --mode quick --no-commit. If the flag is
absent because of version skew, run the command without it, expect the scaffold
to auto-commit, record that generated commit, and land the wrapper artifacts in
a follow-up commit (Orc W1 scaffold evidence).
- Verify scaffold substitution AND advance the lifecycle in
state.md.
On oat ≥0.1.65, verify that the basic scaffold placeholders
({ OAT_HILL_CHECKPOINTS }, { OAT_PHASE }, { OAT_WORKFLOW_MODE }) were
already substituted; fix them only if unexpected survivors remain. The
lifecycle advance and wave-specific values remain orchestrator-owned:
oat_dispatch_policy (managed/high unless the operator says otherwise),
associated_issues (the wave's backlog items),
oat_parallel_execution: true, and — once plan.md is written —
oat_phase: plan + oat_phase_status: complete. Also refresh the
state.md BODY prose (Status/Current Phase/Artifacts/Progress/Next Milestone)
— frontmatter alone is not enough; two waves hit body-drift findings.
- Write
plan.md from assets/wrapper-plan-template.md — pointer-only tasks,
the wrapper execution contract verbatim, drift-refresh record marked
non-authoritative recon evidence, HiLL at the final phase (confirm against
workflow.hillCheckpointDefault). Template directive hygiene: the
templates' { curly-brace } placeholders AND their authoring directives
(e.g. "Keep both code + artifact rows below") are instantiation instructions —
substitute/apply them and REMOVE them from the instantiated artifact; a
surviving directive is reviewer-visible noise (sol wave-2 finding class).
- Write
discovery.md (inherited contract + this wave's decisions) and
orchestration-log.md from assets/orchestration-log-template.md (day one).
- [JUDGMENT] Compose parallel groups from the recon write surfaces: groups of
≤ ceiling, write-disjoint within a group; shared-file plans in different groups.
A lane that must run alone (merges-first, or a solo finale) stays UNGROUPED —
ungrouped phases execute sequentially in plan order, and
validate-plan
REJECTS singleton groups. Cross-model review steps on
locking/security/containment/dependency lanes. Recon must intersect ALL
per-plan write surfaces mechanically, not just headline files (wave-3 g1
missed a two-lane CLI-file overlap that merged conflict-free only by region
luck).
oat project validate-plan --project-path <path>, run the repo's formatter
over the project dir (source-program example: pnpm format:fix), and commit
the scaffold.
Step 4: Plan gate
Run the cross-runtime artifact gate with a bounded prompt (rule 6): review the
wrapper artifacts for plan invariants, contract consistency, frontmatter validity,
and whether any task restates/narrows its source plan — the external plans are
immutable inputs, NOT review targets. Disposition findings in-artifact
(gate-invoked artifact review) and commit. A plan gate MAY PROCEED at
fixes_completed per the wave-0/1 precedent, but that is a proceed point, not a
terminal state. Every gate row MUST flip to passed once all fix dispositions
carry the stored verification records required by the fix-disposition contract below;
passed is the only
terminal state for gate rows (Orc operator-audit S8).
Step 5: Execute via oat-project-implement
The lifecycle skill owns execution. This skill contributes the templates it uses:
- Worktree bootstrap: run
scripts/bootstrap-group.sh <wave-prefix> <BASE_SHA> pNN... — e.g.
scripts/bootstrap-group.sh wave-2 $(git rev-parse HEAD) p01 p02 p03
(wraps the oat-worktree-bootstrap-auto contract: create at explicit base,
propagate local config, verify base ancestry, repo bootstrap and
proportionate baseline via the OAT_WAVE_BOOTSTRAP_CMD /
OAT_WAVE_BASELINE_CMD env hooks — a pnpm-shaped repo defaults to
pnpm run worktree:init / pnpm type-check, any other repo skips the step
with a STATUS line — verify provider-view parity with the root checkout,
structured STATUS lines; as of 1.3.0 the script relocates its
.bootstrap-*.log files into $TMPDIR itself).
When lanes carry provider review steps that require worktree trust, pre-trust
the new paths using that provider's configuration (for Codex, the
source-program example is ~/.codex/config.toml).
- Implementer briefs: self-contained Phase Scope (resolver-stamped dispatch
fields), the contract pointer ("your ENTIRE contract is ; nothing
in the wrapper narrows it"), region-level expected-churn pre-declaration
(rule 4: name the churned file AND the region vs the lane's target region —
three-for-three effective in wave-2 group 2), env rules, DoD gates before
commit, one commit per task with the plan's message, bundle STOP semantics
where applicable (
DR-260713-bundle-stop-semantics-park), the bookkeeping
boundary (workers never touch .oat/projects/), and the structured PHASE
REPORT format. Cross-model review steps in briefs name the RUNTIME-RESOLVED
reviewer (the plan stays provider-neutral; dispatch resolves — currently
Codex per the source repo's config). Adapt lane-type addenda to the consuming
repo. Worked source-program examples: "background a previously-inline step"
lanes budget a full direct-caller audit; new shell scripts honor the repo's
oldest supported shell; disposal/wiring prefers the framework teardown seam
(Fastify onClose in one lane) over a concurrently restructured file; a
defaulted schema field (Zod in one lane) budgets type-surface consumer fixes;
an MCP tool rename moves its expected-tool list, additive snapshot, tool-list
tests, and version-pin tests together; and time-driven HTTP tests avoid a fake
clock that freezes the request event loop. These are briefing patterns, not
requirements for Fastify, vitest, Zod, or MCP.
- Reviewer briefs: read-only, per-phase; checklist = source plan
## Review focus; implementer claims are inputs to VERIFY, not trust.
Lanes with embedded cross-model reviews get DISPOSITION-VERIFICATION briefs:
verify each fix and each rejection's parity claim against the cited siblings
instead of re-reviewing (wave-3 caught a partially-false rejection claim this
way). Every fix disposition, including a root-verified bounded fix, must
produce a minimal stored verification record stating what was verified, how
it was verified, and where the evidence is recorded; the wave-5 final gate
blocked on exactly this audit gap. For a fix continuation, prefer resuming the
original implementer handle while it is alive: this is cheaper and retains
design context. Launch a fresh same-target agent only when that handle is
gone. Rename/refactor-class revision reviews get a PURITY BAR: diff the
changed handler bodies against the ORIGINAL at base — tests alone are silent
on migrated-behavior drift (wave-3 p-rev1 Critical was caught only by this).
LOGIC-BEARING lanes' reviewers must design ≥1 adversarial probe of their own
beyond the implementer's pins (DR-260715-adversarial-probe-reviewer: the
wave-4 overlap-corruption Critical was found ONLY by a reviewer probe —
every gate and pinned test was green); runtime-
probe ambiguous behavioral claims; on containment/security surfaces require a
weaker-anywhere analysis (any input previously rejected that is now accepted is
Critical). Adapt contract-propagation checks to the lane type; the
source-program example required a new DoctorJsonResponse field to travel with
diagnostics docs and exact-match doctor JSON tests. Write the artifact to the
ROOT checkout reviews/ and run the repo's formatter over it (source-program
example: pnpm format:fix).
- Merge choreography: after all group verdicts — serialized
git merge --no-ff in plan order, rebasing each phase branch on the updated tip first
(rules 2–3). Compound the hard pre-merge guard and merge into ONE shell
invocation immediately before EVERY merge:
cd /abs/repo/root && [ "$(git branch --show-current)" = "wave-N-execution" ] || exit 1 && git merge --no-ff ….
The explicit cd repairs the healable cwd dimension; branch drift is
non-healable and hard-aborts. Advisory pwd/branch prints in separate
invocations proved worthless in the Orc W2 incident; this guard prevented two
repeats there, including p10. Integration DoD gates
after fan-in run TO COMPLETION BEFORE any group bookkeeping edits start
(DR-260714-integration-gates-run-before);
then the group bookkeeping commit. Before dispatch, inspect every worktree's
sync commit content and stop on provider-view deletions or unrelated managed
path churn. This is a regression guard for the named stale-local-binary
failure class: a stale locally resolved node_modules/.bin/oat can shadow
the global CLI and make two desired-state sync versions thrash managed files.
On any parity mismatch, compare node_modules/.bin/oat --version with
oat --version; do not treat it as unexplained toolkit corruption.
Conflict-resolution contract
(DR-260715-conflict-resolution-contract): on rebase/merge conflicts —
keep-both where lanes appended to shared surfaces; then in-worktree BUILD +
touched-package suites BEFORE amending (mechanical splices break seams:
braces, stacked branches, duplicated keys — wave-4 evidence); then inspect
the amended commit's file stat against the expected list; NEVER git add -A
in a worktree (stale synced local state gets swept). At every
merge/bookkeeping boundary, run a standing config-integrity check that tracked
.oat/config.json keys remain present; this check originated with
BL-260715-investigate-oat-config-json in the source program's repo. Update
canonical sections in place, the run-entry table, review rows, and state;
remove worktrees + branches after merge.
Step 6: Closeout
Follow oat-project-implement's completion route, plus the wave-specific order.
Run these steps strictly in this numbered sequence. The load-bearing constraint:
the end-of-run synthesis and its summary.md roll-up MUST come before backlog
archival and before the project-archive seal (oat-project-complete) — never
archive anything first.
-
Final verification — integration DoD gates green on the integration branch.
If the repo has no CI, record a one-line explicit waiver in the wave plan:
merge gate = local DoD only. The CI-introducing wave's first green run
certifies the cumulative merged tree and MUST be recorded as closure of that
waiver; do not re-run earlier gates retroactively (Orc W1–W2 evidence).
-
End-of-run synthesis in orchestration-log.md, then roll it up into
summary.md (this is the "before any archive step" gate): convention
verdicts with evidence, adjustments-as-rules for later waves, graduated-entries
ledger, rolled into summary.md ## Workflow Observations.
-
Serialized backlog archival — oat backlog archive with real summaries,
one commit.
-
Root final review.
-
Cross-runtime final gate — judgment-sweep dispositions; after every fix
disposition has its required stored verification record, flip the row to
passed. A final
gate MUST NOT remain at fixes_completed: passed is the only terminal
state for gate rows (Orc operator-audit S8; confirmed convergently by stoa
W6's final-row handling). The historical row-stomp restore-watch is RETIRED:
the upstream stomp class was fixed in oat 0.1.65 and stoa's W6 supplied the
final clean observation (three gate rounds, zero stomps, watch never fired,
2026-07-20).
-
Pre-approval sequence per workflow.postImplementSequence, then a single
HiLL. File follow-up-ledger backlog items at closeout (on main post-merge, or
pre-gate if the operator prefers them in the PR).
-
The full oat-project-complete PROCESS, with an explicit autonomous
deferral branch. Interactive runs retain the standing per-wave order
review → complete → merge (an open PR is expected, not a blocker — the
archive-aware PR body sync handles it). The requirement remains the whole
completion process, named explicitly: oat project complete-state →
oat project archive (the CLI owns the local archive move, the summary
export, and the S3 sync when s3SyncOnComplete is configured) →
active-project pointer clear → the completion bookkeeping commit. Running
oat project complete-state ALONE does NOT satisfy this step: in the Orc
first run all four wrapper projects were left lifecycle-complete but
unarchived until an operator audit asked (S10).
Under autonomous execution, each wave MUST still run
oat project complete-state and its then-current project bookkeeping. The
archive tail — oat project archive (including configured S3 sync) →
active-project pointer clear → completion bookkeeping commit — MAY be
deferred to the program boundary so the wave can merge and execution can
continue. Record every such choice in the wave ledger exactly as
completion tail: deferred to program close; deferral is an outstanding
disposition, never satisfaction of the full-tail requirement. Interactive
per-wave full-tail completion remains valid.
The interactive completion skill is model-invisible
(disable-model-invocation: true), so an autonomous orchestrator executes
its SKILL.md as a document, resolving its gates from config
(workflow.archiveOnComplete, workflow.createPrOnComplete), until an
oat-project-complete-auto companion ships
(BL-260720-add-oat-project-complete-auto). If the archive tail is deferred,
that execution occurs after the one human-gated program-end checkpoint in
oat-wave-program, across every deferred wave wrapper.
-
After the operator merges: reconcile (squash-merge means content-diff the
branch vs main; cherry-pick stragglers), reset the working branch, clean stale
phase branches, and run oat-wave-program wave-close <wave-id> so the
program ledger records the merge (PR, SHA, completion-record link) and flips
the wave's plan rows to done. Per-wave recaps are default-OFF: run one only
on explicit operator request; otherwise record
recap: deferred to program close in the wave ledger. When this wave
completes the final pending wave, offer or run the program recap using the
mechanical explainer caller below. The program recap is generated from the
reconciled program artifact and ALL wave records.
Program-close recap explainer caller
The orchestrator owns fact-base synthesis. At program close it synthesizes an
explainer-kit.fact-base/v1 document from the reconciled execution-program
artifact, ALL wave summaries, and ALL completion records. Its required keys are
exactly:
schemaVersion, generatedAt, mode, freshnessPolicy, sources, claims, unresolvedClaims, overrides.
The caller also owns CONTENT AUTHORING, exactly as it owns critic execution and
fact-base synthesis: the kit's pipeline validates structure and fact
consistency, but nothing in it owns prose quality. An unattended recap run
without a caller-supplied authoring path emits raw federated artifact text as
deck prose (stoa W6 live evidence, run-19af6e55: implementation.md pasted
verbatim, frontmatter included, tables flattened to run-on prose — every
automated gate passed it). The explainer-kit now enforces this seam: every
unattended run requires exactly one provider-neutral author seam — in-process
callers supply an author(request) callback; JSON/CLI callers supply
authorModulePath naming a module with an author function export. The core
invokes it once per recipe artifact with an explainer-kit.author-request/v1
document and expects an explainer-kit.author-result/v1 reply; runs fail if
the author is absent, returns an invalid result, or copies excessive verbatim
source text. Recap callers MUST satisfy that seam by authoring content from
the synthesized fact base plus the recipe outline (LLM-authored from
summary/synthesis material, as the operator-approved W6 rebuild demonstrates)
or NOT run the unattended build, recording the skip disposition per the
optional-step rule. Callbacks and module paths never enter the persisted run
request.
The mechanical caller constructs an explainer-kit.run-request/v1 document whose
required keys are exactly:
schemaVersion, recipe, slug, outputRoot, factBase, mode. Set recipe to
{ "id": "program-recap", "version": "1" }; this is an object with exactly
id and version. Set outputRoot to .oat/repo/explainers/<slug>/. Bind the
synthesized fact-base file through factBase with the required keys
mode, freshnessPolicy, set mode to "supplied", set freshnessPolicy to
"live-wins", and set path to that file.
After the run, read the explainer-kit.manifest/v1 document. Its required keys
are exactly:
schemaVersion, runId, slug, recipe, createdAt, source, theme, artifacts, immutableHashes, outcome, buildRecord, warnings.
Record the default program recap's manifest runId and outcome in the
program ledger; use a wave ledger row only for an explicitly requested per-wave
recap. Publishing is human-gated; this caller never invokes publish.
Success Criteria
- Zero convention re-derivation errors (branch naming, merge types, SHAs, gate
scoping — the wave-1 incident class).
- Every lane: drift check honored, one verified commit per task, review round
dispositioned, DoD + integration gates green.
- Orchestration log current at every group boundary; synthesis written before
archive; follow-up ledger graduated.
- Deviations from THIS skill logged with a skill-abstraction tag — that is the
evidence signal for future skill revisions.